Material identification sensor integrating large language model and deterministic classifier

Abstract

The invention discloses a material recognition sensor integrating a large language model and a deterministic classifier. The material recognition sensor comprises a sensing unit, a signal acquisition unit, a fuzzy logic conversion unit, a data transmission unit, the large language model, the deterministic classifier and a comprehensive recognition unit, the sensing unit senses various non-physical parameters of an object; the signal acquisition unit is used for extracting characteristic data about materials; the fuzzy logic conversion unit converts the data into a fuzzy natural language which can be understood by a large language model; the data transmission unit transmits the natural language generated by the fuzzy logic conversion unit to a large language model; the big language model discriminates the material of the object according to the transmitted natural language to obtain an identification result 1; the deterministic classifier obtains an identification result 2 according to the feature data; the comprehensive identification unit is used for carrying out credibility integration to obtain a final identification result; according to the method, the feature data of the object is converted into the cue words recognized by the large language model, so that material discrimination and application are realized.

Description

Technical Field

[0001] The present invention relates to the field of sensor technology, and in particular to a material recognition sensor integrating a large language model and a deterministic classifier. Background Art

[0002] Large language models (LLMs) are based on natural language text processing, using large amounts of text data and trained through a self-attention transformer mechanism. These models can understand and generate natural language, and reason about human needs and knowledge. However, current large language models (LLMs) cannot directly process sensor data because the sensor signal pattern is very different from text data.

[0003] In industrial production, there are many types of existing material recognition sensors with different signal patterns, which usually require dedicated networks for training. These sensors mostly use methods such as SVM to identify materials in existing databases. For example, the patent application with publication number CN108564094B discloses a material recognition method based on a combination of convolutional neural networks and classifiers. The convolutional neural network model is trained to extract CNN features of the image, and then the extracted CNN features are used to construct a training set. The training set is used to train multiple basic classifiers, and the accuracy of material recognition is improved by combining these basic classifiers. However, there are the following limitations:

[0004] (1) Existing material recognition methods can only classify objects based on existing training samples and cannot identify object types that do not appear in the training samples;

[0005] (2) Material recognition methods usually rely on single parameter measurement, which makes it difficult to accurately identify similar objects;

[0006] (3) Existing sensors can only provide specific parameters and cannot make judgments and interact according to human needs;

[0007] (4) Sensor data is difficult to combine with large models based on natural language. Summary of the invention

[0008] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a material recognition sensor that integrates a large language model and a deterministic classifier, which generates feature data by collecting the mechanical, thermal and electromagnetic parameters of the object, and converts these data into prompt words that can be recognized by the large language model, thereby realizing material discrimination and application.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] A material recognition sensor integrating a large language model and a deterministic classifier, comprising a sensing unit, a signal acquisition unit, a fuzzy logic conversion unit, a data transmission unit, a large language model, a deterministic classifier and a comprehensive recognition unit;

[0011] The sensing unit includes various mechanical, thermal and electromagnetic sensors to sense various physical parameters of the object;

[0012] The signal acquisition unit is used to collect the signal of the sensor unit and perform preprocessing to extract characteristic data about the material;

[0013] The fuzzy logic conversion unit is used to convert the extracted feature data of the material into a fuzzy natural language that can be understood by the large language model;

[0014] The data transmission unit is used to transmit the natural language generated by the fuzzy logic conversion unit to the large language model;

[0015] The large language model determines the material of the object according to the transmitted natural language and obtains recognition result 1;

[0016] The deterministic classifier is used to obtain the recognition result 2 according to the characteristic data of the material;

[0017] The comprehensive recognition unit is used to generate probability distribution functions M1 and M2 of recognition result 1 and recognition result 2, and obtain corresponding credibility by calculating the trust function and likelihood function of the two, so as to perform credibility synthesis and obtain the final recognition result.

[0018] The signal acquisition unit includes two major functions: data processing and feature extraction. The data processing is used to process the time series signal obtained by the sensor unit, including period division of the time series signal obtained by the sensor unit, sliding filtering of abnormal values ​​according to the 3σ principle, and normalization of the acquired mechanical, thermal, and electromagnetic parameters; the feature extraction is used to highlight the characteristic information in the time series signal. For mechanical sensors, the inverse of the resistance is obtained to obtain the contact pressure, which further indicates the hardness of the object; for thermal sensors, the thermal conductivity characteristics of the object are extracted by calculating the amplitude of the temperature signal drop and its time; for electromagnetic sensors, its relatively stable peak value is the relative dielectric constant characteristic.

[0019] The fuzzy logic conversion unit is based on existing samples and uses a triangular membership function to extract the membership of the sample to be tested to the known sample, that is, the fuzzy features of the object. Among the extracted fuzzy features, the two features with the largest membership and the smallest one are selected in terms of mechanics, thermodynamics and electromagnetic features to generate a natural language fuzzy description of the object to be tested, expressing several objects that are most similar to the above-mentioned properties of the object to be tested.

[0020] The data transmission unit can use two modes: wireless transmission or wired transmission. The wireless transmission adopts Bluetooth communication to transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model interface in the computer to realize communication with the large language model; the wired transmission adopts direct connection of data line to directly transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model in the computer.

[0021] The large language model is based on machine learning and deep learning algorithms, and is trained using as many mechanical, thermodynamic and electromagnetic parameters of common materials as possible. These parameters are input into the input layer of the large language model, and matrix operations and nonlinear operations are performed in the hidden layer. The hidden layer parameters are self-learned through the gradient descent method, and finally the material inference results based on extensive knowledge are output in the output layer, and these results are used as the initial membership 1 transmitted to the comprehensive recognition unit.

[0022] The deterministic classifier is a CNN, GRU or LSTM neural network model. The features initially extracted by the signal acquisition unit are input into the input layer of the neural network model of the deterministic classifier. The inner layer of the neural network constructs a function about the initial features through self-learning. The constructed function is used as a further extracted feature to enter the output layer. The output layer obtains the membership of the material to be tested to each training material type by performing matrix operations on the further extracted features. The material with the highest and lowest membership for the mechanical, thermodynamic and electromagnetic features are respectively taken as the initial membership 2 transmitted to the comprehensive recognition unit.

[0023] The specific identification process of the comprehensive identification unit is as follows:

[0024] (1) Generation of probability distribution function: For the initial membership 1 and 2 output by the large language model and the deterministic classifier, the membership of each predicted material is used as the initial value of the corresponding single element in the probability distribution function M1 and M2. For the case where the deterministic classifier does not contain the element with the highest membership in the prediction result of the large language model, the equivalent inference coefficient λ is defined, and the element with the highest membership in the prediction result of the large language model is added to M2, and its initial value is assigned to λ. For the probability distribution assignment of the element combination, the minimum value of the probability distribution function of each element in the combination is used as the probability distribution function assignment of the combination. Finally, M1 and M2 are normalized respectively so that the sum of the probability distribution function values ​​of each subset is 1.

[0025] (2) The calculation of the trust function and likelihood function is as follows:

[0026] The likelihood function is used to calculate the confidence that each subset in the probability distribution function M1 and M2 is true. According to the likelihood function definition:

[0027]

[0028] Among them, Pl() is the likelihood function, Bel() is the trust function, which is defined as:

[0029]

[0030] For the case where A is a single element, Bel(A) = M({A})

[0031] According to the above formula, the likelihood function values ​​Pl1 and Pl2 of each element in the probability distribution function M1 and M2 can be calculated respectively. According to the likelihood function definition, we have:

[0032] CF i (H)=Pl i (H),i=1,2

[0033] Therefore, the confidence CF1 and CF2 of each element in the probability distribution function M1 and M2 can be calculated according to the above formula. Since CF1(H) and CF2(H) are both greater than or equal to 0, the confidence synthesis algorithm gives:

[0034] CF 1,2 (H)=CF1(H)+CF2(H)-CF1(H)·CF2(H)

[0035] Thus, the final confidence of each material is obtained, and the element with the highest final confidence is selected as the final result of material identification.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The fuzzy logic conversion unit of the present invention uses a triangular membership function to extract the membership of the sample to be tested to the known sample, which can extract the fuzzy features of the material and then output its natural language fuzzy description, effectively realizing the combination of material recognition and a large language model.

[0038] 2. The large language model of the present invention is based on machine learning and deep learning algorithms, and is trained using as many mechanical, thermodynamic and electromagnetic parameters of common materials as possible, so that material recognition is trained with a trusted network database, and its recognition has good versatility, professionalism and accuracy.

[0039] 3. The comprehensive recognition unit of the present invention can integrate the recognition results of the large language model and the deterministic classifier with a scientific algorithm, integrating the breadth of the large language model recognition and the accuracy of the deterministic classifier recognition, thereby making the recognition results of the present invention have higher accuracy.

[0040] In summary, compared with the existing technology, the present invention generates feature data by collecting the mechanical, thermal and electromagnetic parameters of the object, and converts these data into prompt words that can be recognized by the large language model. The large language model is based on machine learning and deep learning algorithms, thereby realizing accurate identification and application of materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a structural schematic diagram of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0043] Reference Figure 1 , a material recognition sensor integrating a large language model and a deterministic classifier, comprising a sensing unit, a signal acquisition unit, a fuzzy logic conversion unit, a data transmission unit, a large language model, a deterministic classifier and a comprehensive recognition unit;

[0044] The sensing unit includes various mechanical, thermal and electromagnetic sensors to sense various physical parameters of the object;

[0045] The signal acquisition unit is used to collect the signal of the sensor unit and perform preprocessing to extract characteristic data about the material;

[0046] The fuzzy logic conversion unit is used to convert the extracted feature data of the material into a fuzzy natural language that can be understood by the large language model;

[0047] The data transmission unit is used to transmit the natural language generated by the fuzzy logic conversion unit to the large language model;

[0048] The large language model can identify the material of the object according to the transmitted natural language and obtain the recognition result 1;

[0049] The deterministic classifier is used to obtain the recognition result 2 according to the characteristic data of the material;

[0050] The comprehensive recognition unit is used to generate probability distribution functions M1 and M2 of recognition result 1 and recognition result 2, and obtain corresponding credibility by calculating the trust function and likelihood function of the two, so as to perform credibility synthesis and obtain the final recognition result.

[0051] The sensing unit comprises:

[0052] Mechanical sensor: used to measure the contact force between the sensor and the material, which is subsequently converted into a hardness index by the neural network;

[0053] Thermal sensor: Contains a ring resistor. When powered on, the voltage across the ring resistor is stable, and the sensor has a constant temperature when thermal equilibrium is reached. After contacting an object, since the object is at room temperature and the temperature is relatively low, heat is transferred from the sensor to the object, and the thermal conductivity of the object can be calculated from the rate at which the object's temperature rises.

[0054] Electromagnetic sensor: used to measure the capacitance between the sensor and the object, and then use the formula Calculate the relative dielectric constant ε of the object.

[0055] The signal acquisition unit includes two major functions: data processing and feature extraction. The data processing is used to process the timing signal obtained by the sensing unit, including periodic division of the timing signal obtained by the sensing unit (one period is from the sensor contacting the object to leaving the object), sliding filtering of abnormal values ​​according to the 3σ principle, and normalization of the obtained mechanical, thermal, and electromagnetic parameters; feature extraction is used to highlight the characteristic information in the timing signal. For mechanical sensors, since the signal obtained is actually the size of the resistor in the circuit, and the size of the resistor is inversely proportional to the size of the contact pressure, it is necessary to find the inverse of the size of the resistor to obtain the contact pressure, which further represents the hardness of the object; for thermal sensors, the thermal conductivity characteristics of the object are extracted by calculating the amplitude of the temperature signal drop and its time; for electromagnetic sensors, its more stable peak value is the relative dielectric constant characteristic.

[0056] The fuzzy logic conversion unit uses a triangular membership function based on existing samples to extract the membership of the sample to be tested to the known sample, that is, the fuzzy features of the object. Among the extracted fuzzy features, two features with the largest membership and one feature with the smallest membership are selected in terms of mechanics, thermodynamics and electromagnetic features to generate a natural language fuzzy description of the object to be tested, describing several objects that are most similar to the above properties of the object to be tested.

[0057] For example, the relative dielectric constant of object A' measured by the present invention is ε A' , then calculate its membership to the existing samples A, B, C, D, E. Assume that the relative dielectric constant of object A is ε A , define the membership function of object A' to object A in terms of relative dielectric constant as a triangular membership function, that is, In ε A' =ε A When ε , the function reaches the maximum value of 1, indicating that at this time, the membership of object A' to object A in terms of relative dielectric constant is 1, that is, at this time, object A' is almost consistent with object A in terms of relative dielectric constant; A' = 0 or 2ε A or ε A' >2ε A(From physics knowledge, we know that the relative dielectric constant of any object is greater than or equal to 0), the function takes the minimum value 0, indicating that the relative dielectric constant of object A' is completely different from that of object A. Similarly, the membership of object A' to objects B, C, D, and E can be calculated. The membership μ calculated is fuzzy graded. A For example, 0.9≤μ A When ≤1, object A' is said to be "particularly similar" to object A in terms of relative dielectric constant; 0.8≤μ A <0.9, object A' is said to be "relatively similar" to object A in terms of relative dielectric constant; 0.6≤μ A <0.8, object A' is said to be "somewhat similar" to object A in terms of relative dielectric constant; 0.5≤μ A <0.6, object A' is said to be "almost dissimilar" to object A in terms of relative permittivity; 0.3≤μ A <0.5, object A' is said to be "not very similar" to object A in terms of relative dielectric constant; 0≤μ A <0.3, object A' is said to be "dissimilar" to object A in terms of relative permittivity. Take the two objects with the largest and smallest membership in terms of relative permittivity of object A', assuming they are C and E, respectively, and μ C =0.96,μ E = 0.35. Repeat similar operations for hardness parameters and thermal conductivity parameters.

[0058] The information obtained by the feature fuzzy module is converted into natural language that can be understood by the large language model. For example, for μ C =0.96,μ E = 0.35, the natural language generation module converts it into “Object A' is very similar to object C in terms of relative permittivity, and almost dissimilar to object E.”

[0059] The data transmission unit can use two modes: wireless transmission and wired transmission. The wireless transmission adopts Bluetooth communication to transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model interface in the computer to realize communication with the large language model; the wired transmission adopts direct connection through a data cable to directly transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model in the computer.

[0060] The large language model is based on machine learning and deep learning algorithms, and is trained using the mechanical, thermodynamic and electromagnetic parameters of nearly all common materials. These parameters are input into the input layer of the large language model, and complex matrix operations and nonlinear operations are performed in the hidden layer. The hidden layer parameters are self-learned through the gradient descent method, and finally the material inference results based on extensive knowledge are output in the output layer, and these results are used as the initial membership 1 transmitted to the comprehensive recognition unit.

[0061] The deterministic classifier is a neural network model such as CNN, GRU, LSTM, etc. The features initially extracted by the signal acquisition unit are input into the input layer of these neural network models. The inner layer of the neural network constructs functions about the initial features through self-learning. These constructed functions are used as further extracted features to enter the output layer. The output layer obtains the membership of the material to be tested to each training material type by performing matrix operations on the further extracted features. For the mechanical, thermodynamic and electromagnetic features, the material with the highest and lowest membership is taken respectively as the initial membership 2 transmitted to the comprehensive recognition unit.

[0062] The comprehensive identification unit includes: generation of probability distribution function, calculation of trust function and likelihood function and confidence synthesis algorithm.

[0063] Generation of probability allocation function: For the initial membership 1 and 2 output by the large language model and deterministic classifier, the membership of each predicted material is used as the initial value of the corresponding single element in the probability allocation function M1 and M2. For the case where the deterministic classifier does not contain the element with the highest membership in the prediction result of the large language model, define the equivalent inference coefficient λ, and add the element with the highest membership in the prediction result of the large language model to M2, and assign its initial value to λ. For the probability allocation assignment of element combinations, use the minimum value of the probability allocation function of each element in the combination as the probability allocation function assignment of the combination. Finally, M1 and M2 are normalized so that the sum of the probability allocation function values ​​of their subsets is 1.

[0064] For the recognition result 1 of the large language model, assume that the large language model infers that the probability of object A' being metal gold, metal copper and glass is 0.95, 0.6 and 0.2 respectively. Then we calculate the individual values ​​and combined values ​​in the probability distribution function in turn, and finally normalize them. That is, we already have M1({gold})=0.95, M1({copper})=0.6, M1({glass})=0.2; then calculate the probability of the combination, that is, M1({gold, copper})=min{M1({gold}),M1({copper})}=0.6,

[0065] M1({gold, glass})=min{M1({gold}),M1({glass})}=0.2,

[0066] M1({copper, glass})=min{M1({copper}),M1({glass})}=0.2,

[0067] M1({gold, copper, glass})=min{M1({gold}),M1({copper}),M1({glass})}

[0068] =0.2,

[0069] Then normalize it, we have:

[0070]

[0071] Since 0.273+0.242×2+0.061×4=1, no correction is needed. Otherwise, M1({gold, copper, glass}) should be corrected according to M1({gold})=1-[M1({gold})+M1({copper})+M1({glass})+M1({gold, copper})+M1({gold, glass})+M1({copper, glass})].

[0072] For the recognition result 2 of the deterministic classifier, assume that the probability of inferring that object A' is copper and glass (assuming that no sample of metal gold has been input into the deterministic classifier at this time) is 0.6 and 0.2 respectively. The probability that the deterministic classifier infers that object A' is gold is called the equivalent inference coefficient of object A', represented by λ. A more conservative assumption is made here, so that the deterministic classifier neither infers that object A' is gold nor denies that object A' is gold, that is, λ=0.5, so when normalization is not performed, M2({gold})=0.5. By the same token, we can get

[0073] M2({copper})=0.250,M2({gold})=0.208,

[0074] M2({glass})=0.083,M2({copper, gold})=0.250,

[0075] M2({copper, glass})=0.083,M2({gold, glass})=0.083,

[0076] M2({copper, gold, glass})=0.085.

[0077] Calculation of trust function and likelihood function:

[0078] The likelihood function is used to calculate the confidence that each subset in the probability distribution function M1 and M2 is true. According to the likelihood function definition:

[0079]

[0080] Where Pl() is the likelihood function. Bel() is the trust function, which is defined as:

[0081]

[0082] For the case where A is a single element, Bel(A) = M({A})

[0083] According to the above formula, the likelihood function values ​​Pl1 and Pl2 of each element in M1 and M2 can be calculated respectively. According to the likelihood function definition, we have:

[0084] CF i (H)=Pl i (H),i=1,2

[0085] Thus, the confidence CF1 and CF2 of each element in M1 and M2 can be calculated according to the above formula. Since CF1(H) and CF2(H) in the application of the present invention are both greater than or equal to 0, the confidence synthesis algorithm has:

[0086] CF 1,2 (H)=CF1(H)+CF2(H)-CF1(H)·CF2(H)

[0087] Thus, the final confidence of each material can be obtained. The element with the highest final confidence is selected as the final result of material identification. For the above case, there are:

[0088] Similarly

[0089] Confidence synthesis algorithm:

[0090] Since CF1(H0) and CF2(H1) in the present invention are both greater than or equal to 0, according to the confidence synthesis algorithm,

[0091] CF 1,2 (H)=CF1(H)+CF2(H)-CF1(H)·CF2(H)

[0092] Thus, the final confidence of each material can be obtained. The element with the highest final confidence is selected as the final result of material recognition.

[0093] That is, it is known that CF1({gold})=pl1({gold})=0.661,

[0094] CF1({copper})=Pl1({copper})=0.542,

[0095] CF1({glass})=P11({glass})=0.272;

[0096] CF2({gold})=Pl2({gold})=0.584,

[0097] CF2({copper})=Pl2({copper})=0.626,

[0098] CF2({glass})=Pl2({glass})=0.334,

[0099] According to the uncertainty synthesis algorithm, we have:

[0100]

[0101] In the application of the present invention, CF1(H) and CF2(H) are both greater than or equal to 0, so:

[0102] CF 1,2 (H)=CF1(H)+CF2(H)-CF1(H)CF2(H)

[0103] Therefore, CF 1,2 (gold)=CF1(gold)+CF2(gold)-CF1(gold)CF2(gold)=0.859

[0104] Similarly, CF 1,2 (Copper) = 0.829, CF 1,2 (Glass) = 0.515

[0105] Therefore, CF 1,2 (Gold)>CF 1,2 (Copper)>CF 1,2 (Glass)

[0106] Therefore, it is finally inferred that object A' is metal gold.

[0107] In summary, the sensor of the present invention successfully identified a material type that is outside the training sample of the deterministic classifier, which shows that the present invention has a wide range of recognition; if the large model initially infers M1({gold})=0.7, M1({copper})=0.6, M1({glass})=0.2, then CF 1,2 (gold)=0.846 <CF 1,2 (copper) = 0.848, indicating that when the certainty of using the large language model for material inference is not high, the prediction result of the large language model will not be easily accepted, which shows the rigor and accuracy of the present invention; when the equivalent inference coefficient λ defined in the present invention is changed, the conservatism of material identification can be changed. The larger λ is, the more radical the result of material identification tends to be; the smaller λ is, the more conservative the result of material identification tends to be, which shows the flexibility of the present invention in actual use.

[0108] Example: Application of the present invention to mine exploration

[0109] In the exploration of unknown mines, it is first necessary to determine the chemical composition of the mine environment to avoid possible safety risks. Since different substances may have similar colors and lighting equipment may not be suitable for unknown environments, the visual recognition method is less effective. At this time, the material recognition sensor combined with the large language model of the present invention can be considered. The present invention can be installed on a mechanical gripper to infer the types and uses of substances contained in the mine environment, and then infer their chemical properties.

[0110] In this embodiment, the sensing unit uses pressure-sensitive materials and interdigital electrode sensors to measure the force between the grasped object and the mechanical gripper during the grasping process; a thin film heating wire component is used to measure the thermal conductivity of the grasped object (the temperature of the contacted object only increases by 5-10°C, which will not cause safety hazards); and an interdigital electrode is used to measure the dielectric constant of the grasped object to obtain mechanical, thermal and electromagnetic parameters. The signal acquisition unit is connected to the sensing unit, collects the signal generated by the sensing unit, performs pre-processing such as filtering on the collected signal, and extracts the characteristic data about the material. The fuzzy logic conversion unit converts these characteristic data into natural language fuzzy descriptions, such as "the hardness of the object is approximately 70% of copper, the relative dielectric constant is similar to that of gold, and the thermal conductivity is similar to that of glass." The data transmission unit uses Bluetooth communication to transmit the fuzzy natural language from the fuzzy logic conversion unit to the large language model. The large language model is based on transformer and is pre-trained using a large amount of natural language information. The large language model receives information from the data transmission unit and performs intelligent processing, outputs the judgment of the material and the possible use of the material, and obtains the recognition result 1. Result 1 includes possible classifications of objects and the confidence levels of different classifications. After receiving result 1, the comprehensive recognition unit normalizes the confidence levels in result 1 to obtain a probability distribution function M1. At the same time, the feature data of the material extracted by the signal acquisition unit is also transmitted to the deterministic classifier via Bluetooth communication. The deterministic classifier outputs the probability that the material belongs to each training category, and the comprehensive recognition unit calculates the probability distribution function M2. The comprehensive recognition unit calculates the trust functions of M1 and M2, and then obtains their likelihood functions. The likelihood function is used as the inferred credibility CF1 and CF2 of the two, thereby obtaining the comprehensive credibility CF 1,2 Finally, the comprehensive credibility CF is taken 1,2 The largest element is taken as the final recognition result.

Claims

1. A material recognition sensor integrating a large language model and a deterministic classifier, characterized in that: It includes a sensing unit, a signal acquisition unit, a fuzzy logic conversion unit, a data transmission unit, a large language model, a deterministic classifier, and a comprehensive recognition unit; The sensing unit includes various mechanical, thermal and electromagnetic sensors to sense various physical parameters of the object; The signal acquisition unit is used to collect the signal of the sensor unit and perform preprocessing to extract feature data about the material; The fuzzy logic conversion unit is used to convert the extracted feature data of the material into a fuzzy natural language that can be understood by the large language model; The data transmission unit is used to transmit the natural language generated by the fuzzy logic conversion unit to the large language model; The large language model can identify the material of the object according to the transmitted natural language and obtain the recognition result 1; The deterministic classifier is used to obtain the recognition result 2 according to the characteristic data of the material; The comprehensive recognition unit is used to generate probability distribution functions M1 and M2 of recognition result 1 and recognition result 2, and obtain corresponding credibility by calculating the trust function and likelihood function of the two, so as to perform credibility synthesis and obtain the final recognition result.

2. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The signal acquisition unit includes two major functions: data processing and feature extraction. The data processing is used to process the time series signal obtained by the sensor unit, including period division of the time series signal obtained by the sensor unit, sliding filtering of abnormal values ​​according to the 3σ principle, and normalization of the acquired mechanical, thermal, and electromagnetic parameters; the feature extraction is used to highlight the characteristic information in the time series signal. For mechanical sensors, the inverse of the resistance is obtained to obtain the contact pressure, which further indicates the hardness of the object; for thermal sensors, the thermal conductivity characteristics of the object are extracted by calculating the amplitude of the temperature signal drop and its time; for electromagnetic sensors, its relatively stable peak value is the relative dielectric constant characteristic.

3. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The fuzzy logic conversion unit is based on existing samples and uses a triangular membership function to extract the membership of the sample to be tested to the known sample, that is, the fuzzy features of the object. Among the extracted fuzzy features, the two features with the largest membership and the smallest one are selected in terms of mechanics, thermodynamics and electromagnetic features to generate a natural language fuzzy description of the object to be tested, expressing several objects that are most similar to the above-mentioned properties of the object to be tested.

4. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The data transmission unit can use two modes: wireless transmission or wired transmission. The wireless transmission adopts Bluetooth communication to transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model interface in the computer to realize communication with the large language model; the wired transmission adopts direct connection via data cable to directly transmit the natural language fuzzy description generated by the fuzzy logic conversion unit to the large language model in the computer.

5. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The large language model is based on machine learning and deep learning algorithms, and is trained using as many mechanical, thermodynamic and electromagnetic parameters of common materials as possible. These parameters are input into the input layer of the large language model, and matrix operations and nonlinear operations are performed in the hidden layer. The hidden layer parameters are self-learned through the gradient descent method, and finally the material inference results based on extensive knowledge are output in the output layer, and these results are used as the initial membership 1 transmitted to the comprehensive recognition unit.

6. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The deterministic classifier is a CNN, GRU or LSTM neural network model. The features initially extracted by the signal acquisition unit are input into the input layer of the neural network model of the deterministic classifier. The inner layer of the neural network constructs a function about the initial features through self-learning. The constructed function is used as a further extracted feature to enter the output layer. The output layer obtains the membership of the material to be tested to each training material type by performing matrix operations on the further extracted features. The material with the highest and lowest membership for the mechanical, thermodynamic and electromagnetic features are respectively taken as the initial membership 2 transmitted to the comprehensive recognition unit.

7. The material recognition sensor integrating a large language model and a deterministic classifier according to claim 1, characterized in that: The specific identification process of the comprehensive identification unit is as follows: (1) Generation of probability distribution function: For the initial membership 1 and 2 output by the large language model and the deterministic classifier, the membership of each predicted material is used as the initial value of the corresponding single element in the probability distribution function M1 and M2. For the case where the deterministic classifier does not contain the element with the highest membership in the prediction result of the large language model, the equivalent inference coefficient λ is defined, and the element with the highest membership in the prediction result of the large language model is added to M2, and its initial value is assigned to λ. For the probability distribution assignment of the element combination, the minimum value of the probability distribution function of each element in the combination is used as the probability distribution function assignment of the combination. Finally, M1 and M2 are normalized respectively so that the sum of the probability distribution function values ​​of each subset is 1. (2) The calculation of the trust function and likelihood function is as follows: The likelihood function is used to calculate the confidence that each subset in the probability distribution function M1 and M2 is true. According to the likelihood function definition: Among them, Pl() is the likelihood function, Bel() is the trust function, which is defined as: For the case where A is a single element, Bel(A)=M({A}) According to the above formula, the likelihood function values ​​Pl1 and Pl2 of each element in M1 and M2 can be calculated respectively. According to the likelihood function definition, we have: CF i (H)=Pl i (H),i=1,2 Therefore, the confidence CF1 and CF2 of each element in M1 and M2 can be calculated according to the above formula. Since CF1(H) and CF2(H) are both greater than or equal to 0, the confidence synthesis algorithm gives: CF 1,2 (H)=CF1(H)+CF2(H)-CF1(H)·CF2(H) Thus, the final confidence of each material is obtained, and the element with the highest final confidence is selected as the final result of material identification.

Images